在SVM中使用特征之前如何在特征上使用L2规范化

Question

I'm training SVM with features extracted from Convolutional Neural Network. As written in this paper ( http://arxiv.org/pdf/1405.3531v4.pdf ) it is good to L2-normalize your features before applying SVM on them.

I use this function to normalize vectors:

def l2normalize(features):
    l2norm = np.sum(np.abs(features)**2,axis=-1)**(1./2)
    return features/l2norm

After this normalization my accuracy drops from around 60% to 20% so something is clearly wrong. How should I properly prepare my vectors for SVM with L2 norm?

Answer 1

1. Check if you are actually normalizing both the features extracted from training data and the test data.

2. Try using normalize from sklearn.preprocessing . I checked and your function returns slightly different values.